Deepak Jose

Design, Ceo-office Strategic Projects at @RushGamingu with @team_hike, Ex-Head of Product Design and User Research, @Dunzoit | Ex @Olacabs @hikeapp

[Deepak Jose @iDeepakJose]

@EleanorKonik @gauravvohra I have same set up.

[Deepak Jose @iDeepakJose]

@pmitu 💯

[Deepak Jose @iDeepakJose]

@kavinbm Inspiring @kavinbm 🫡. Lia’s desk is such a cool idea.

[Deepak Jose @iDeepakJose]

Insane! This is a peek into how you build world-class product with ai🫡

[Kavin @kavinbm]

24 days. Just me + Claude Code + Agents. No co-founder. No team. Here's the damage report building Lia (will post optimized Claude.md soon).

Feb 20 - Ground Zero
• Woke up and decided my AI should manage my life, not just answer questions
• MVP live by end of day
• Gmail, Calendar, Drive, voice, memory, all wired
• Intelligent model routing from day 1
• First session felt like talking to a new hire who already read every email I've ever sent

Feb 21
• Needed a front door. Built one
• getlia dot ai. Live, animated, slightly too pretty for a product no one can use yet

Feb 22
• Before connecting to anyone's inbox. 4 layers of security
• Pattern matching, content sandboxing, ML-based scanning, full audit trail
• Wordmark done. She has a name now

Feb 23
• Killed the pricing tiers. One plan. If you need a Chief of Staff, you need the whole thing
• Referral system live. Your position in line moves when you bring friends
• Started modeling unit economics. Margins look… interesting

Feb 25
• RICE scored 38 features. Ruthless prioritization
• Designed a prompt architecture that lets the AI load context on demand instead of stuffing everything into one prompt
• 10 reference files specced for dynamic loading
• Locked the voice. She sounds like her now
Feb 26
• Rewired how the AI thinks. Two-tier reference system, dynamic loading
• Context engine design finalized
• The boring infrastructure that makes the magic feel effortless

Feb 27
• WhatsApp backend integration done. Telegram first, WhatsApp next
• Built a cost model that actually tells me whether this business works at scale
• Modeled cache strategies, rate limits, per-user economics
• It works. Barely. With the right model routing

Feb 28
• The big one. 28 commits in a single day
• Memory upgrade. Hot memory (markdown) + cold memory (Postgres + pgvector). She remembers everything now. Not just the last conversation. Everything. Huge upgrade from OpenClaw
• Context engine live: ingestion, memory, enrichment, delivery.
• Relationship scoring. Pattern detection. Style correction
• DB-backed session persistence. Conversations survive deploys and restarts
• Full architecture docs written
• This is the day Lia stopped being a chatbot

Mar 1
• Trust system shipped. Lia earns autonomy over time
• Day 1 she asks permission for everything. Day 30 she just handles it
• Trust decays if you stop using a feature. She doesn't assume
• Confirmation gate on every destructive action. She never sends without asking
• LLM reliability framework. Hallucination checks, conflation defense, failure mode catalog
• Session persistence. Full tool blocks saved to DB. She remembers drafted emails and pending confirmations across restarts
• VIP alerts v2. Smart summaries with Read/Draft/Snooze buttons
• Replaced the entire Google integration layer. Direct API calls, no CLI dependency

Mar 2
• Calendar holds. Propose meeting times, hold them on your calendar, auto-release when someone picks
• Email follow-up tracking v1. Thread-level detection, enrichment-based priority
• Memory upgrade again. Search, threads, compaction, confidence scores
• Draft narration verify layer. Auto-retry when the model describes changes instead of actually rewriting
• Feedback pipeline. User reports → GitHub Issue → Claude Code analyzes → status tracking
• Admin dashboard. Trust scores, feedback, waitlist, retention stats. Full operational visibility
• Trial wind-down and churn handling. 3 final messages, subscribe button, hard block, returning user flow
• Re-engagement system. Day 7 check-in, Day 40 subscribe-or-delete

Mar 3
• Smart Opus routing. Sonnet by default, classifier upgrades to Opus for compose, sensitive, complex tasks
• Prompt improvements. Temperature routing, anti-hallucination guards, XML structure. Voice trimmed 22%
• Lia learns your email writing style per recipient. Tone, greetings, sign-offs. Drafts sound like you, not a bot
• Reply-all default, thread auto-read, email signature extraction and injection
• Conversational feedback bridge. Trust adjustments from memory corrections

Mar 4
• Waitlist live. Substack + X + LinkedIn. Same morning
• Scheduling Inbound V1. Someone emails wanting to meet. Lia reads it, proposes timezone-aware slots, drafts replies in your voice, books with Meet link
• WhatsApp v2. Full feature parity with Telegram
• Web search via Perplexity Sonar
• Admin dashboard dark mode
• Centralized error messages across 9 tool surfaces. No more LLM improvising technical jargon to users
• 7 new test commands for internal QA

Mar 5
• Test coverage day. Not glamorous. Necessary
• Eval pass rate: 31.7% → 90.9% across 11 test suites
• 36 E2E workflow tests. 122 unit test files. 576 unit tests
• CI coverage gates. E2 90%, E3 90%. Tests can't silently regress
• Drafts no longer disappear on failed send. Auto-save to Gmail Drafts
• The kind of day that makes the next 100 days possible

Mar 6
• 13 features moved to shipped in one sweep
• Email Triage V2. Every email scored. Morning briefs ranked by what actually matters
• VIP emails that aren't urgent get batched for morning. No more false alarms
• Self-hosted website analytics. Page views, scroll depth, CTA clicks. Zero cookies
• Admin dashboard waitlist funnel. Full conversion tracking from visit to card reserved
• She now asks "what would you like me to remember?" instead of blindly saving
• Proactively catches you up when you return after being away
• 8 behavioral fixes. The ones where she gets noticeably smarter overnight

Mar 7
• Morning Brief V2. One message at 8am. Everything you need. Action loop walks you through items one by one
• Evening Wrap-Up at 6pm. What happened, what's open, what's tomorrow
• Self-healing error loop. Sentry catches it, Claude Code fixes it, PR opens, CI passes, auto-merges. I wake up to a fix
• 10x faster morning briefings. Parallel processing
• Meeting logistics. Flags missing video links before your meeting starts. One-tap Add Google Meet
• Lia ramps up. Plugs into your tools and gets to work on your world during onboarding
• Sentry SDK across all 4 Railway services
• API semaphore + retry on rate limits. No more dropped messages under load
• 576 unit tests. 90% coverage on critical paths
• Structured JSON logging across all services
• Real relationship scoring wired into email triage. Weekly refresh

Mar 9
• Voice overhaul. ElevenLabs Scribe v2 + Eleven v3. She sounds better and hears better
• Killed Google Gemini from the stack entirely
• Official spend tracking. I know exactly what each user costs me, per provider, per day
• Thread-aware VIP alerts. She knows when a VIP is replying to something you talked about
• Document summarization. Attachment downloads. She reads PDFs, Word docs, spreadsheets
• Prompt caching + Claude 4.6 = ~65% cost reduction stacked
• Context compaction. Conversations never hit the context window ceiling
• Scheduling lookahead extended to 60 days
• Happiness score per user on admin dashboard
• She detects when you're frustrated and backs off. When you're happy, she leans in
• Quiet hours. No proactive messages during sleep
• Orchestrator refactor. Shared handler for Telegram + WhatsApp

Mar 10
• Weekend mode. Saturday briefings shift to calendar-first, no inbox anxiety energy on Sundays
• Users can kill weekend briefings entirely with one sentence
• Morning briefings now auto-advance through items and wrap up. No more "Next?" prompts
• Email confirmations use real names instead of raw addresses
• Scheduling respects the exact duration you ask for. No more silent 30-minute defaults
• Unified all database user IDs to bigint. The boring migration that prevents a catastrophic one later

Mar 11
• E2E test suite fully migrated to cloud. All 39 workflows run in GitHub Actions against production
• No local bot connection needed. No local OAuth tokens. Everything tests against the real thing
• 3 new E2E workflows: email search, weekend briefing, web search
• E2E report redesigned as a chat-style thread with clickable status icons
• Test tiers documented: Tier 1 (every deploy), Tier 2 (nightly), Tier 3 (weekly)
• The test infrastructure is now more reliable than most startups' production infrastructure

Mar 13
• Lia's Desk shipped. Telegram Mini App
• One screen. Everything she's tracking. Urgent emails, overdue commitments, outbound emails awaiting reply, upcoming calendar
• Outbound email tracking. Lia detects your sent emails that haven't gotten a reply and surfaces them
• Smart filtering. "Thanks!" and "Sounds good" emails don't trigger false waiting cards
• Not a task manager. A confidence board. Glance, feel good, close

Mar 14
• Watchtower. Reliability monitoring across all 4 services
• Sentry issue poller. Every 5 minutes, checks for new errors. Auto-triggers Claude Code to diagnose and create fix PRs
• Autofix PRs are PM-friendly: What broke, Why, What this fixes, Risk assessment. Never auto-merges
• When a fix PR gets merged, the Sentry issue auto-resolves. The loop closes itself
• Credit balance now lives in DB with a dashboard button. No more stale env vars
• Google connect webview finally auto-dismisses on both iOS and Android after OAuth
• Timezone detection no longer breaks when the AI wraps its response in markdown
• 18 things shipped in one day. A Wednesday

Mar 15
• Quality Observatory. A monitoring dashboard for output quality. Self-healing error loop. Reliability monitoring. The system watches itself now.
• User trust scores, rejection trends, acceptance rates by output type, proactive engagement rates
• 30-day rejection trend chart. Click through to any user's detail
• Quality alerts on Telegram. 5 conditions: important user frustrated, trust drop, rejection spike, output type failing, proactive fatigue
• Three alert levels: silent, inform, act
• Quality monitor runs every 5 minutes inside the health loop
• Onboarding email briefing no longer sends twice when generation is slow
• 18 new tests for the quality system alone

Lia is going live soon. Waiting for Google Cloud to verify the app ⚡️

[Anish Moonka @anishmoonka]

https://t.co/up3tP8ers1

[Deepak Jose @iDeepakJose]

💯

[Uncle J @UncleJAI]

@figma The loop is tightening.

Code → Design → Feedback → Code

The bottleneck is no longer execution.
It’s taste + constraints.

[Deepak Jose @iDeepakJose]

I’m truly humbled by all the messages from everyone who reached out after my podcast with Darshan Suthar — thank you! 🙏

🎥 If you haven’t caught it yet, here’s the link to watch: https://t.co/Bop2ra9D88

[Deepak Jose @iDeepakJose]

Huge thanks to Darshan Suthar for hosting me!

🎙️ Check out the goldmine of videos in his channel — packed with insights.

https://t.co/SxYT7Txa2m

[Deepak Jose @iDeepakJose]

🎥 I recently shared these lessons in a talk — from “wow moments” to creative trust.
If you’ve ever wondered how to build products people truly love, this one’s for you:
https://t.co/Bop2ra95iA

[Deepak Jose @iDeepakJose]

Grateful for the journey so far, and excited for the chapters ahead — Going OffBarrier 🚀

[Deepak Jose @iDeepakJose]

How We Built Products Millions Couldn’t Stop Loving ❤️
Grateful to have built first-of-its-kind products with amazing teams — embraced by millions of Indians far beyond what we imagined.

[Deepak Jose @iDeepakJose]

🎥 I recently shared these lessons in a talk — from “wow moments” to creative trust.
If you’ve ever wondered how to build products people truly love, this one’s for you:
https://t.co/Bop2ra95iA

[Mohit Kumar @mohitkumar0708]

About last night

@iDeepakJose @agniva_si @iamarnob6543 @ashwinikumrr

[Deepak Jose @iDeepakJose]

Had a great time & wonderful chat with everyone 😀

[Mohit Kumar @mohitkumar0708]

Had a wonderful time reconnecting with folks from my Ola days!

@iDeepakJose @agniva_si @Ranook_xd

[Deepak Jose @iDeepakJose]

🔥

[Kavin @kavinbm]

The dream team @Team_Hike

[Pauca @paucastudio]

New template drop!

Meet 𝗢𝗯𝗿𝗮 — A design studio template for @Framer.

It will be 𝗙𝗥𝗘𝗘 for the next 𝟮𝟰𝗵!

Follow, retweet and comment 𝗢𝗯𝗿𝗮 and I’ll send u the remix link.

Preview: https://t.co/Kg13WomaZd
Buy: https://t.co/USU1FX1eRk

[Deepak Jose @iDeepakJose]

@kanso__cc @framer Obra

[Deepak Jose @iDeepakJose]

Incredible

[Sam Altman @sama]

too long for a tweet, some thoughts on GPT-4o:

https://t.co/6FIdBUn539